draft not for citation - african development bank€¦ · draft not for citation ... 1 the first...
TRANSCRIPT
Exploring differences in national and international poverty estimates: Is Uganda on track to halve poverty by 2015?
Sebastian Levine
This version: 13 October 2010
DRAFT NOT FOR CITATION
United Nations Development Programme Plot 24 Prince Charles Drive
Kampala, Uganda +256‐312‐338100
ABSTRACT
This paper explores differences in the estimates of incidence of poverty in Uganda since the early 1990s as measured by the national poverty line used by Uganda Bureau of Statistics and the World Bank’s “dollar a day” international poverty line. While both measures point to a declining trend in poverty over time there are important differences in the levels of poverty, the speed of the decline and the direction of change in the late 1990s. Proximate causes of this divergence are related to differences in the specification of mean level of welfare and in the distribution of welfare. Proposed underlying causes include differences in: adjustments to account for urban and rural price variation, adjustments to reflect household composition and the application of probability weights associated with the sampling structure.
Keywords: Uganda, sub‐Saharan Africa, Poverty line, Purchasing Power Parities, Millennium Development Goals
The views expressed in the paper are those of the author and do not represent UNDP. I am grateful to Jean‐Yves Duclos and Abdelkrim Araar for insightful comments. I am also grateful for the support from James Muwonge and Sarah Ssewanyana in accessing the national data and for many discussions on the measurement of poverty in Uganda. Omissions and commissions are my own.
- 1 -
1. Introduction
Ten years after the Millennium Development Goals (MDGs) were conceived, and
with just five years to their 2015 deadline, there are still important conceptual and
methodological challenges outstanding on how to assess progress towards the MDG
targets.1 This is especially true for the first and probably most prominent target: “To
reduce by half the proportion of people living on less than a dollar a day.”2 In this
paper, these challenges are discussed in the context of Uganda; a country often
highlighted for being one of a few in sub‐Saharan Africa on track to meet the poverty
target of the MDGs, and for its effective development planning framework and its
strong poverty monitoring system.
When comparing the estimates of poverty incidence from the World Bank’s
interactive database PovcalNet with those from the Uganda Bureau of Statistics
(UBOS) contained in the most recent official MDG progress report, three distinct
discrepancies emerge. Firstly, the estimates generated by the World Bank are much
higher than the national estimates. Secondly, while both sets of estimates show a
declining trend in poverty levels since the early 1990s, the national poverty
estimates point to a rate of poverty reduction that is much faster than when using
the international estimates. The implication is that, when looking at national
poverty estimates, Uganda appears to be on track to meet MDG1/1 and even the
Government’s own more ambitious targets. However, when using international
estimates the country is clearly not on track. The third deviation is that according to
the national estimates from UBOS there was a significant increase in poverty levels
between 1999 and 2002, which is not reflected in the international estimates from
the World Bank.
In light of this ambiguous evidence the central question remains: is Uganda
on track to meet the first MDG and halve poverty by 2015? To seek answers to that
question this paper explores methodological and conceptual differences in the way
the poverty measures for Uganda are being computed and reported by the national
1 The first set of Millennium Development Goals, targets and indicators were outlined in United Nations (2001). For an updated list see http://mdgs.un.org. 2 Since this is the first target of the first Goal below it shall be referred to as “MDG1/1”.
- 2 -
authorities and the World Bank. The analysis should have considerable relevance
for how the Government, domestic stakeholders and international development
partners assess the success of the Uganda Poverty Eradication Action Plan—
especially its promises of “pro‐poor growth”—and commence implementation of
the new National Development Plan in 2010. Moreover, there are important
implications for how national governments and international institutions
collaborate around measuring development progress, which are particularly
pertinent as the global community looks towards a post‐2015 agenda.
The paper is organised as follows. Section 2 discusses poverty trends
according to a range of data sources, but the focus is mainly on the estimates
presented in the national MDG Report, that represent the official estimates of
poverty from the Government of Uganda, and the estimates contained in PovcalNet,
which contains the most recent update of the World Bank’s estimates for poverty.
Section 3 introduces a simple diagnostic framework for exploring the possible
causes of diverging poverty estimates. In Section 4 this framework is used to explore
the differences in the estimates of poverty in Uganda from the World Bank and
national authorities but it should be general enough to guide a comparison of
poverty estimates whether for one country over several time periods or between
different countries.
On this basis, it is possible to explain most of the divergence through
differences in the way price changes and differences are dealt with, the application
of probability weights in the sample survey data and especially how adjustments are
made for household composition. However, given limitations in the meta‐data for
the poverty estimates produced by the World Bank these results are necessarily
based on interpretation of—and some degree of speculation on—the incomplete
information that is made available, and even then a part of the divergence in the
estimates cannot be directly accounted for. Accordingly, Section 5 concludes and
offers some recommendations for how to improve the information base for national
and international poverty estimates.
- 3 -
2. Progress towards MDG 1/1 in Uganda
Uganda is often regarded as a best‐practice or even a “show case” when it comes to
reducing poverty, fighting HIV/AIDS and promoting economic and social
development (McGee 2004; Dijkstra and van Donge 2001). Per capita GDP has
grown 3 percent on average in real terms since the early 1990s outpacing most
countries in sub‐Saharan Africa. As a result GDP per capita has grown more than 60
percent from USD 175 in 1991 to 282 in 2007. Economic growth was particularly
pro‐poor in the first part of the period and is thus key to explaining the fast rates of
poverty reduction (Kappel et al. 2005). Uganda’s strong record in reducing poverty
since the early 1990’s also means that its country report on progress towards the
MDGs is one of the few in sub‐Saharan Africa that rates as “probable” the likelihood
of reaching MDG1/1. Indeed according to the UN Economic Commission for Africa,
Uganda is among just eight countries out of 44 in sub‐Saharan Africa that are likely
to reach the MDG poverty target.3
Moreover, the earlier country’s development planning framework, the
Uganda Poverty Eradication Action Plan (PEAP), with its emphasis on national
ownership and participatory processes, became the model for the Poverty
Reduction Strategy Papers introduced by the World Bank and the IMF in 1999 to
assist Highly Indebted Poor Countries (HIPC) towards identifying their development
challenges and guide them towards debt‐relief (Mallaby 2004; Mackinnon and
Reinikka 2002). As a consequence, Uganda was also the first country to qualify for
debt relief under the HIPC initiative.
2.1 Monitoring at national, regional and global levels
The MDGs constitute the quantitative and time‐bound goals that were designed as
part of the follow‐up to the Millennium Declaration of 2000. On the global level,
international organisations track progress according to their thematic mandate, e.g.
World Bank on the poverty goals of MDG1, UNESCO on the education goals of MDG2
3 The others are: Botswana, Burkina Faso, Cameroon, Ghana, Lesotho, Mauritius and South Africa (http://www.uneca.org/mdgs/MDGs_page.asp accessed June 2010).
- 4 -
and so on. Cross‐sectoral databases administered by the UN Statistics Division and
reports such as the “Millennium Development Goals Report” from the UN4 and the
joint “Global Monitoring Reports” of the World Bank and IMF5 provide regular
updates of global and regional progress against all the goals and targets. Regional
reports are produced by regional development banks, the economic commissions of
the UN and other entities. At country‐level, national governments and their
development partners, including the UN agencies, are expected to prepare national
MDG reports. Hundreds of such “MDGRs” have been prepared by both developed
and developing countries, with several countries having gone through more than
one reporting cycle. In many countries civil‐society organisations, with support
from the Millennium Campaign, also prepare “shadow reports” that provide an
alternative, and sometimes more critical, assessment of country progress.
In Uganda the national authorities and the United Nations Country Team
have collaborated on two national reports to assess progress the country’s towards
the MDGs in 2005 and 2007 (an updated report is due for release in 2010). These
reports follow standard guidelines issued by the United Nations Development Group
(2003) and thus using national data sources, the reports go one by one in describing
status and trends for each goal and assessing the likelihood of their attainment.
However, because of differences in the methodologies used by international
agencies for purposes of global reporting on one hand, and by national authorities
assessing progress locally on the other hand, it is only natural that discrepancies
between the various types of estimates will arise.
Such challenges are particularly pertinent when it comes to target 1a: “Halve,
between 1990 and 2015, the proportion of people whose income is less than one
dollar a day.” The main issue here is that “less than one dollar a day” refers directly
to the international poverty line used by the World Bank, which is generally
considered of limited practical use for national governments and statistics offices.
As a result, the UN Statistics Division in its official guide to monitoring progress
towards the MDGs advises that two indicators should be used: (A) Proportion of
4 See: http://www.un.org/millenniumgoals/ 5 See: http://go.worldbank.org/UVQMEYED00
- 5 -
population below $1 (PPP) per day and (B) Poverty headcount ratio (% of
population below the national poverty line), and it is explicitly stated that: “For
monitoring country poverty trends, indicators based on national poverty lines should
be used, where available” (United Nations 2003: xi).
For reporting purposes, when possible, the World Bank has started listing
the two types of indicators side by side in its flagship World Development Report
and the UN Statistics Division includes both measures on its website dedicated to
monitoring progress towards the MDGs. For purposes of monitoring MDG1/1 a
division of labour as such has thus been established whereby World Bank issues its
estimates of national, regional and global levels of poverty using the US$1/day
poverty line, and national statistics offices issue their estimates using the national
poverty lines. But what do we do when this division of labour generates
contradictory evidence on the poverty trajectory? The remainder of this paper
discusses this challenge.
2.2 Overview of the poverty estimates in Uganda
The two main sources that track progress towards MDG1/1 in Uganda are reflected
on Figure 1. The graph shows the incidence of poverty for Uganda according to the
national MDG report from 2007 (UN 2007) and the updated figures from the World
Bank’s PovcalNet.6 According to the first source, which is based on estimates by
national authorities in UBOS (2006), the incidence of poverty has been reduced
from 55.7 percent in 1992/1993 to 31.1 percent in 2005/2006. The long‐term
decline in the incidence of poverty was briefly interrupted between 1999/2000 and
2002/2003 when the incidence increased from 33.8 to 37.7 percent. Nevertheless,
over the period poverty reduction has proceeded with great speed and fast enough
for the country—assuming a linear extrapolation of current trends—to be on track
to meet the nationally defined target under MDG1/1, which is set to half the value of
the incidence observed in 1992/1993, or 28 percent, by 2015. The slightly more 6 The World Bank requests that all outputs from PovcalNet should cite the source as “PovcalNet: the on‐line tool for poverty measurement developed by the Development Research Group of the World Bank” and provide the web link to PovcalNet: http://iresearch.worldbank.org/PovcalNet/.
- 6 -
ambitious target set out in the new National Development Plan that began
implementation in 2010, of reducing poverty to 24.5 percent by 2014/2015 also
seems within reach. In fact, progress has been so fast that the country appears to be
on track to reach the even more ambitious goal of the first PEAP in 1997, which was
to reduce poverty incidence to 10 percent in 2017. It is also this earlier target that is
reported in the MDG report.
Figure 1 also includes the poverty incidence estimates released by the World
Bank based on its international “dollar a day” poverty line, which is expressed in
US$ Purchasing Power Parties (PPP) to facilitate international comparison (more on
this below). This time series begins in 1989 but also shows a downward trend from
1992 as the national poverty estimates. However, it is also evident that there are
some important differences. Firstly, the World Bank estimates of poverty are much
higher: 64 higher percent to be precise using the most recent figures. Secondly,
while the overall trend is down for both measures the progress in poverty reduction
is slower under the World Bank estimates. In fact, extrapolating the past rate of
progress Uganda is slightly off‐track in terms of meeting the MDG target that is
derived as a proportion of poverty incidence in 1992. Using World Bank estimates
the PEAP target also appears unattainable. Thirdly, the increase in poverty observed
when using the UBOS estimates in the middle of the period is not evident from the
PovcalNet data as poverty incidence falls from 60.5 percent to 57.4 percent between
1999/2000 and 2002/2003.
More details on the poverty incidence data is provided in Table 1, which
includes the figures from a variety of sources that have computed estimates based
on the nine household surveys that UBOS has conducted since the late 1980s. The
characteristics of the different surveys, the Household Budget Survey (HBS), the
Integrated Household Survey (HIS), the Monitoring Surveys (MS) and the Uganda
National Household Surveys (UNHS) and issues related to survey design and
execution will be discussed in more detail below. In row 1‐12 on Table 1 are listed
the studies undertaken by national authorities or researchers working with local
institutions. Most of these studies present own estimates while a few, including the
MDG report (row 11), base their estimates on the works of others (mostly on UBOS
- 7 -
1999, 2003 and 2006). Small discrepancies in the early studies (row 1‐3) appear to
be a result of continuous cleaning of the raw data (Appleton 1998). However,
studies conducted after the 2002/2003 survey had to take into consideration the
inaccessibility of certain parts of the country due to insecurity, an issue also
discussed further below.
Figure 1: Trends and targets of poverty incidence in Uganda
Source: UBOS (2006), Ministry of Finance, Planning and Economic Development (2004), UN (2007) and PovcalNet
(Accessed June 2010).
In the past, World Bank estimates for Uganda were reported less regularly
and prior to the update of PovcalNet in October 2008 the most recent World Bank
estimate that appears to be for 1992 as reported in the World Development Report
(WDR) 2000/2001 (row 16). The two earlier figures are reported for 1989 are 50
percent (row 13) and 69.3 percent (row 14‐15) although the former is supposed to
cover an extended period of time between 1981 and 1995. Since the late 1990’s the
WDR has reported side‐by‐side the poverty estimates generated by the World Bank
and those generated by national authorities. However, for Uganda this has
55.7(92/93)
45.0%(97)
33.8(99/00)
37.7%(02/03)
31.1%(05/06)
MDG target 2015: 28%
PEAP(I) target for 2017: 10%
68.7%(89/90)
70.0%(92/93)
64.4% (95/96) 60.5%
(99/00) 57.4%(02/03)
51.5%(05/06)
NDP targetfor 2014/5: 24.5%
MDG target2015: 35%
0%
10%
20%
30%
40%
50%
60%
70%
80%
1985 1990 1995 2000 2005 2010 2015 2020
Start of Poverty Eradication Action Plan (PEAP)
World Bank/PovcalNet (US$1.25/day in 2005 PPP)
UBOS (Ush 21135.37 in 1997 prices)
Start of National Development Plan (NDP)
- 8 -
happened only twice, in WDRs of 1999/2000 and 2000/2001. Moreover, the WDR
of 2002 reported the national estimate for 1999/2000 but omits data for this year in
the five subsequent editions of the report.
Table 1: Estimates of poverty incidence in Uganda (%)
89/90 92/93 93/94 94/95 95/96 97 99/00 02/03 05/06
HBS HIS MS-1 MS-2 MS-3 MS-4 UNHS-1 UNHS-2 UNHS-3
UBOS affiliated research
1. Appleton (1998) 55.6 50.3 49.2 45.6
2. Appleton (1999) 55.5 52.5 50.1 48.5 44
3. Appleton (2003) 55.7 51.2 50.2 49.1 44.4 35.1
4. UBOS (2003) 55.7 44.4 38.8
5. UBOS (2003) 33.8/1 37.7/1
6. Appleton and Ssewanyana (2003) 55.7 44.4/1 33.8 37.7/1
7. Okidi et al. (2004) 55.7 45 33.8/1 37.7/1
8. Kappel et al. (2005) 55.7 33.8/1 37.7/1
9. UBOS (2005) 56.4 38.8 31.1
10. UBOS (2005) 55.7/1 33.8/1 37.7/1 28.9/1
11. MDGR (United Nations 2007) 55.7 45 33.8/1 37.7/1 31.1
12. Ssewanyana and Okidi (2007) 56.4 38.8 31.1
Mean (1-12) 55.81 51.33 49.83 47.73 44.53 34.02 38.11 31.38
Standard deviation (1-12) 0.318 1.106 0.551 1.872 0.393 0.531 0.569 0.550
World Bank publications
13. WDR97 50.0/2,3
14. WDR98/99 69.3/2
15. WDR99/00 69.3/2 55.0
16. WDR00/01 55.0/36.7/2
17. WDR02 35.2
18. WDR03 55.0
19. WDR04
20. WDR05 44.0
21. WDR06 55.0 44.0
22. WDR07 33.8 37.7
23. WDR08 (World Bank 2008b) 33.8 37.7
24. World Bank (2006) 55.7 37.7
25. Unstat.un.org (Accessed Jun 10) 33.8 37.7
26. PovcalNET (Accessed Jun 10) 68.7/2 70.1/2 64.4/2 60.5/2 57.4/2 51.5/2
Notes: 1/ The source explicitly states that the estimate is based on a data sample that excludes Kitgum, Gulu, Bundibugyo, Kasese and Pader districts 2/ Uses international poverty line expressed in US$ PPP 3/ Period covers 1981-95
Sources: As indicated and various annual editions of the World Bank’s World Development Report (WDR).
- 9 -
Other international sources including the Poverty and Vulnerability
Assessment by World Bank (2006), the WDR of 2008 and the website of the UN
Statistics Division all use the national estimates generated by UNOS. It is also clear
from the table that there is some divergence in which surveys are included in the
various reports. Early national studies (Appleton 1998, 1999, 2003) report
incidence for the Monitoring Surveys on the 1990’s while more recent studies omit
these (UBOS 2005, Ssewanyana and Okidi 2007) altogether or include only MS‐4
(UBOS 2003, Appleton and Ssewanyana 2003, UN 2007). PovcalNet on the other
hand includes MS‐3 and along with the early WDRs it is the only source also to
include the HBS of 1989/1990.
This brief overview of poverty estimates from the national and international
sources suggests a number of discrepancies notably in terms of level of poverty
incidence and the degree of poverty reduction over the period, which leaves open
the question about whether Uganda is on track to meet the MDG1/1 target.
3. Diagnosing divergence in poverty estimates
This section presents a diagnostic framework for analysing the differences between
estimates of poverty. The framework is applied to explore the differences between
national and international estimates of poverty in Uganda but it should be general
enough to guide a comparison of any two (or more) sets of poverty estimates
whether for one country over several time periods or between different countries.
3.1 A diagnostic framework
In the proposed diagnostic framework a distinction is made between “proximate”
and “ultimate” causes of the divergence in poverty estimates. For understanding
proximate causes it is useful to think of poverty measures as being determined by
three elements: the relative inequalities in the measure of welfare, differences in the
mean value of that measure, and differences in the cut‐off point that separates the
poor from the non‐poor, again using the same measure of welfare. Any difference
- 10 -
between two poverty estimates for two distributions A and B can therefore be
written as:
∆ ; ; ; ;
where P is the poverty measure, z is the poverty line, μ is the mean of the welfare
measure and L represents the Lorenz‐curve. From the online version of PovcalNet it
is possible to get information about all three determinants of P and these can be
compared with the micro‐data that is made available by UBOS to provide proximate
causes of ∆ .
Figure 2: Comparison of normalised welfare measures
Sources: Author’s calculations based on data from PovcalNet and UBOS.
Note: Welfare measures are normalised by the respective poverty lines.
Figure 2 compares the mean levels of the welfare measures (μ) used by UBOS
and PovcalNet normalised by their respective poverty lines (z). There are some
noteworthy results. Firstly, normalised mean levels of welfare for UBOS estimates
were higher than PovcalNet throughout the period. Secondly, the difference was
growing over the period: in 1992/1993 the difference between the two sets of
normalised welfare measures was 31 percent rising to 63 percent in 2005/2006.
Thirdly, changes were uneven over the four surveys. While the normalised welfare
measure drawn from PovcalNet increased by 19 percent between 1992/1993 and
1999/2000 the corresponding increase was 48 percent in the estimates from UBOS.
- 11 -
However, in the next period the picture was reversed. Between the surveys in
1999/2000 and 2002/2003 the normalised welfare measure for PovcalNet
increased by 12 percent, while the UBOS measure increased by 5 percent. These
results point to a much faster reduction in poverty using the national data on mean
welfare and the national poverty line. But this will also depend on differences in the
relative inequalities.
Figure 3: Differences in cumulative shares of welfare by ranked quinquintile (PovcalNetUBOS)
Sources: Author’s calculations based on data from PovcalNet and UBOS.
Differences in relative incomes can be gauged on Figure 3, which shows the
differences in the cumulative measure of welfare at quin‐quintiles (5 percent) for
the two distributions. These differences are all negative, indicating that the Lorenz
curve for PovcalNet estimates are lower than UBOS estimates for every survey‐to‐
survey comparison. In other words, the welfare distribution used in PovcalNet
displays a greater degree of inequality (in technical terms, the UBOS estimates of
inequality second order dominate the PovcalNet estimates). Whereas the
- 12 -
differences in mean levels of welfare seemed to have increased over time, the
differences between the two sets of estimates in terms of the relative inequalities
appear to have narrowed over the course of the four surveys. Clearly, this is not an
indication that inequality has fallen (in fact it has increased as will be shown below),
but just that the differences in inequality between the two sources of the poverty
estimates have reduced.
This initial investigation into the proximate causes suggest that the
differences in poverty measures are a result of differences in the mean level of
welfare and/or the poverty line, as well as the relative inequalities in the
distribution of welfare. With lower levels of normalised mean welfare levels and a
greater degree of inequality it is not a surprise that PovcalNet generates higher
estimates of poverty.
Figure 4: Causes of differences in poverty estimates
Poverty estimate, P
A. Lorenz curve, L(p)
B. Mean welfare, μ
C. Poverty line, z
1. Data organization and aggregation/ calculation method (+)
2. Construction of poverty line (+) 3. Value of poverty line/ method for updating (+)
4. Indicator of welfare/ definition (‐/+)
Survey instrument/data Adjustments 5. Unit of analysis (‐) 11. Prices (+)6. Survey structure (+) 12. Household composition (+) 7. Sampling errors (+) 13. Household size (‐)8. Non‐sampling errors (+) 14. National accounts (‐)
9. Missing values (‐)10. Data sources (+)
Notes: Proximate causes of divergent poverty estimates are numbered alphabetically and underlying causes are numbred numerically. An underlying cause that is likely to impact the comparison of estimates on poverty in Uganda is indicated with (+).Causes that may be important in other types of comparisons are indicated with (-).
It is possible to link these proximate causes with a total of 14 possible
ultimate causes related to a range of methodological issues that affecting directly or
indirectly the main elements of the poverty estimates. These ultimate causes, and
- 13 -
which of the proximate causes they are likely to influence, are presented on Figure
4, which will serve as a checklist for the diagnostic analysis that follows. Ultimate
causes are drawn from the extensive empirical and theoretical literature on poverty
measurement (e.g. Deaton 1998; Datt 1998; Ravallion 1998; Lanjouw and Ravallion
1995; Ravallion and Bidani 1994). Issues related to data organisation and
disaagregation of data affect directly the estimation of the Lorenz curve (item 1 on
Figure 4) as explained further below. Generally, the level of the poverty line will be
directly affected by the way it is constructed and how it is updated (2‐3). A series of
causes are grouped as reflecting differences in the survey instruments and the
underlying data (5‐10), and another series of causes are grouped as reflecting
differences in the way a number of possible data adjustments are being made (11‐
14). These are linked to the estimates of poverty either directly or indirectly
through the indicator of welfare (4) as will be explained further below.
The remainder of this section discusses those potential underlying that can
be ruled out early on as having an impact in the comparison of international and
national poverty estimates in Uganda. The next section is devoted to those issues
that are found likely, on the basis of the limited meta‐data available in PovcalNet, to
have an impact and the magnitude of these is assessed.
3.2 Issues that can be ruled out
A number of ultimate causes can be ruled out fairly quickly in the specific case of
comparing the poverty estimates in Uganda. Both sets of estimates by the World
Bank and UBOS/EPRC in Uganda are based on consumption expenditure data from
household surveys conducted by UBOS over the past several decades. These surveys
are listed on Table 2. Consumption expenditure is collected using recall questions
on a range of common and infrequent food and non‐food expenditure. It is therefore
possible to rule out that the discrepancy in poverty estimates accrues from
differences in the definition of welfare (item 4 in Figure 4) for instance if one source
used income and another using expenditure. However, a number of issues related to
the survey data and survey instrument, and adjustments made to the data, affect the
estimation of the welfare measure causing the differences in the mean and
- 14 -
distribution of the welfare measure discussed above. This is also the main reason
that “Welfare measure/definition” is placed centrally in the graphic presentation of
the diagnostic framework, because it may work on its own or in combination with
other causes to affect the poverty measure. In order words, while there is
consistency in the broad definition of welfare, how that welfare is estimated is a
different matter.
Table 2: The household surveys
Survey Round Dates Sample size
Household Budget Survey (HBS) Apr. 1989 - Mar. 1990 4,595
Integrated household survey (IHS) Mar. 1992 - Mar. 1993 9,925
Monitoring survey 1 (MS-1) Aug. 1993 - Feb. 1994 4,925
Monitoring survey 2 (MS-2) Jul. 1994 - Jan. 1995 4,925
Monitoring survey 3 (MS-3) Aug. 1995 - Jun. 1996 5,515
Monitoring survey 4 (MS-4) Mar. 1997 - Nov. 1997 6,654
Uganda National Household survey 1 (UNHS-1) Aug. 1999 - Jul. 2000 10,696
Uganda National Household survey 2 (UNHS-2) May. 2002 - Apr. 2003 9,711
Uganda National Household survey 3 (UNHS-3) May. 2005 - Apr. 2006 7,417
Source: Ssewanyana and Muwonge (2004) and UBOS (2006).
The next issue that can be ruled out relates to the unit of analysis (item 5),
which for both sets of estimates is the household, as household consumption
expenditure forms the basis for the poverty estimates. When computing measures
of poverty, UBOS and EPRC in Uganda weigh estimates by the size of the household
so that the incidence of poverty is expressed as a percentage of individuals. It is not
immediately clear from the PovcalNet meta‐data how this is done but presumably
this is done before the aggregated data on the welfare distribution and the mean
incomes entered into the database by World Bank country teams. The main
argument for weighting poverty measures by household size is that poor
households are typically larger, so the share of poor individuals exceeds the share of
poor households, and for most types of poverty analysis the interest is primarily in
people as individuals. That is also why MDG1/1 refers to the share of the population
below the poverty line, not the share of households.
- 15 -
The organisation of the data and associated methods is another factor that
can be ruled out fairly quickly (item 1). This issue arises from the fact that PovcalNet
relies on distributional data that is typically available only in grouped form, such as
income shares of deciles of household ranked by per capita income.7 In the national
poverty reports, estimates of poverty incidence is computed directly using unit data
and the formula for the normalised FGT index (after Foster, Greer and Thorbecke,
1984), or:
;∑
∑
where yi is consumption expenditure for household i and yi ≤ z, n is the sample size
and hi and wi are weights that account for the number of household members and
the probability weight of the household that is linked to the sampling structure,
respectively. The poverty aversion parameter, α, takes a value of 0 in case of the
poverty headcount (H). For the poverty gap (PG) then α =1 and for the squared
poverty gap (SPG) then α = 2.
However, when using grouped data the approach is to use parametric
specifications of the underlying Lorenz curve, from which a range of poverty and
inequality measures can be calculated. Two functional forms are used to estimate the
Lorenz curve in PovcalNet referred to as General Quadratic (CQ), and Beta‐Lorenz
curve (Datt 1998). H is then computed using the established relationship between the
inverted Lorenz curve L'(p) and the distribution function, which when evaluated at
the poverty line is: ⁄ . Other poverty measures are given by:
⁄ , and SPG is estimated as the integral of 1 ⁄ over the range
of (0, H).8
In Table 3 the FGT poverty measures and other commonly poverty and welfare
measures are computed using the Uganda household survey data for 1992/1993 and
2005/2006 using the data in grouped form and in unit data form. The grouped form
7 The meta‐data for Uganda in PovcalNet indicates that the micro‐data was available but the computations use the grouped format (http://iresearch.worldbank.org/PovcalNet/doc/UGA.htm#3; accessed June 2010). 8 See Datt (1998) for details on the estimation of these measures under the two Lorenz curve specifications.
- 16 -
data was generated from the household surveys and inputted into the downloadable
version of the Povcal software that allows users to conduct welfare analysis using the
PovcalNet methodology outlined above.9 Data can be entered in a variety of way, but
in this case the inputs were: percentage of the population in a given class interval of incomes and the mean income of that class interval (“type 5” in the Povcal
terminology).
Table 3: Poverty and inequality measures from unit and grouped data
FGT poverty measures:
Fractiles
Best fit Lorenz-
curve Mean (USh)
α=0 α =1 α=2 Watts Gini MLD
1992/1993
Grouped data 5 Beta 23861.58 56.2 20.9 10.3 30.1 35.8 22.2
10 Beta 23861.58 56.2 20.9 10.3 30.1 35.8 22.1
20 Beta 23861.58 56.1 20.9 10.3 30.0 35.7 22.0
Unit data N=9923 23862.48 56.4 20.9 10.3 30.0 35.7 21.6
(22932.71-24792.25)
(54.2-58.6)
(19.7-22.1)
(9.6-11.1)
(28.4-32.6)
(34.2-37.3)
(19.7-23.4)
2005/2006
Grouped data 5 Beta 39472.85 31.1 8.8 3.5 11.3 40.2 27.5
10 Beta 39472.85 31.2 8.9 3.6 11.4 40.1 27.3
20 GQ 39472.85 31.4 8.9 3.3 11.7 39.9 26.6
Unit data N=7421 39469.73 31.1 8.8 3.5 11.5 39.9 26.4
(37743.8-41195.65)
(29.2-33.0)
(8.1-9.4)
(3.2-3.9)
(10.7-12.3)
(38.4-41.4)
(24.4-28.4)
Notes: GQ = General Quadratic Lorenz-curve, Beta = Beta Lorenz-curve. MLD = Mean Log Deviation. Poverty line = USh 21135.17. Means are in 1997 prices. Figures in brackets represent 95% confidence intervals estimated on the unit data using Stata’s svy command and the survey sampling structure.
Source: Author’s computations based on UBOS data for the full sample surveyed.
As a test of robustness grouped data in three different fractiles were tested
for quintiles, deciles and quin‐quintiles and the results are presented for whichever
Lorenz‐curve estimate fitted the data best (generally the two fits were very close).
The unit form data is run using automated routines in STATA10 that allow for
specification of the survey structure and estimates of standard errors.10 From the
presented results it is clear that the using grouped form data provides estimates 9 http://go.worldbank.org/YMRH2NT5V0 10 The analysis in this paper draws on the comprehensive collection of automated routines contained in the Distributive Analysis for STATA Package (DASP) by Araar and Duclos (2006).
- 17 -
that are quite close to the estimates from the unit data and within the margin of
error at conventional levels of confidence. Actually, sampling variability appears to
be more important than the organisation of the data (item 7). This should be
sufficient evidence to rule out the use of grouped form data in PovcalNet as a source
of divergence in poverty estimates in Uganda. This also corresponds with Datt’s own
tests using data on India that support the validity of the parametric approach.
The treatment of missing and zero‐values (item 9) can have great influence
on estimates of poverty incidence (Szekely et al. 2000). However, this tends to
mostly affect income‐based poverty measures as most households are likely to
report some consumption expenditure even if they have no income, e.g. subsistence
farmers. In the 2005/2006 survey UBOS thus dropped just 5 households due to
missing or zero values leading to a reduction in the number of households used in
the analysis from 7,426 to 7,421 (UBOS 2006: 52).11 Given this small number,
differences in treatment is unlikely to lead to significant differences in results.
As part of the poverty analysis stage it is customary to make a series of
adjustments to the data and the poverty line. Two such common adjustments do not
play a role in explaining the divergence in poverty estimates in Uganda. These relate
to adjustments for economies of scale (item 13) and the use of national accounts
multipliers (item 14). There is ample evidence that adjusting for economies of scale
can have large impact on measures of poverty although there is little consensus
especially when it comes to the size of possible scale economies. Often in empirical
studies economies of scale are included as part of sensitivity and robustness tests
(Lanjouw and Ravallion, 1995). Neither the national estimates of poverty nor those
generated by PovcalNet include adjustments for economies of scale.
Moreover, in some countries household expenditure are adjusted using
national accounts data, which is a subject of considerable methodological debate
(e.g Deaton 2001). However, no such adjustments have been conducted on the
Uganda data neither for the MDGR nor the PovcalNet estimates. According to Chen
and Ravallion (2008) consumption estimates from national accounts data are used
11 These numbers differ slightly from those in Table 2 and as shall be explored further there are also some discrepancies when it comes to the sample size figures reported in PovcalNet.
- 18 -
in the global estimates to fill gaps for countries with only few surveys. However,
PovcalNet reports only country estimates for survey years. In other words, national
accounts adjustments seem to affect regional and global estimates of poverty but
not country level estimates of poverty as presented in PovcalNet. Indeed the
PovcalNet metadata states that per capita survey mean at local currency used in the
case of Uganda.
4. Underlying causes for divergence in poverty estimates
This section explores those issues that are likely to play a role on explaining the
divergence in the national and international estimates and seeks to quantify the
impact of these differences. First addressed are issues related to the value and
construction of the two poverty lines, then the impact of differences in adjustments
for household composition, organisation of the data and sampling issues are
discussed, and finally the role of price adjustments.
4.1 The construction and value of the poverty line
Poverty lines can differ in construct in a number of ways and in turn poverty
estimates can be highly sensitive to the specification of the poverty line. Given the
important differences between the poverty lines used in the national and
international estimates it is useful to briefly highlight the methodology underlying
each.
The Ugandan poverty line
The poverty line in Uganda was developed in Appleton (1999) and has since formed
the basis for analysis of income poverty of the household survey data collected by
UBOS. The poverty line follows the Cost of Basic Needs (CBS) approach presented in
Ravallion and Bidani (1994) and consists of a food and a non‐food component. The
food component was originally estimated based on the consumption patterns of the
1993/1994 MS‐1. First a food basket was identified with 28 of the most frequently
consumed food items by households with less than median income. These food
items were then converted into their caloric equivalent and scaled to generate 3,000
- 19 -
calories per adult equivalent day using as reference the WHO estimates for an 18‐30
year old male. Whilst the 3000 calorie per adult equivalent requirement seems
rather high compared to that assumed in poverty lines for other countries, the per
capita requirement is not so far from requirements assumed in other studies
(Appleton, 1999).
The associated food poverty line is national and uniform across geographical
regions. Non‐food requirements are estimated as the non‐food expenditure of those
households whose total expenditure is just equal to the food poverty line. The
resultant poverty line is thus what Ravallion (1998) refers to as the “lower bound”
value in contrast to the “upper bound” poverty line generated through an estimate
of non food expenditure of households whose food expenditure is equal to the food
poverty line. The estimation method for non‐food items allows different locations to
have different non‐food requirements for instance as urban dwellers often have to
pay more for given accommodation than those residing in rural areas. This way the
non‐food component allows for regional variation unlike the food component. As a
result Uganda has nine poverty lines, for urban and rural each of the four
administrative regions (that is, Central, Eastern, Northern and Western), and one
national poverty line (Table 4).
Table 4: Uganda poverty lines (in 1997 Uganda Shilling) per adult equivalent
Region Total Poverty line
Food poverty line
Central rural 21322.23 15327.45
Central urban 23149.64 15327.45
Eastern rural 20651.86 15327.45
Eastern urban 22125.24 15327.45
Northern rural 20871.98 15327.45
Northern urban 21799.82 15327.45
Western rural 20308.17 15327.45
Western urban 21625.72 15327.45
National 21135.37 15327.45
Source: Author’s communication with EPRC.
- 20 -
While the household surveys are conducted by UBOS the poverty estimates are
computed by independent researchers at the Economic Policy Research Center
(EPRC) following the methodology outlined above. The first poverty line was
expressed in 1993 prices and has since been rebased to 1997 prices by using the
national CPI to inflate the food and non‐food components separately.
The International Poverty Line
For the purposes of measuring global poverty the international poverty line of the
World Bank aims to apply a common standard for all countries anchored in the
poverty experiences of the poorest ones. The basic idea is to make sure that two
people who command the same purchasing power over commodities are classified
consistently as either poor or non‐poor irrespective of whether they live in same or
different countries. There are two major methodological challenges involved in
setting the poverty line. The first is to define a common unit of measurement that
allows for comparability across countries. The second is to derive a uniform poverty
threshold for distinguishing the world’s poor from the non‐poor. The way both of
these methodological challenges are resolved have implications for the
comparability of the international and national poverty lines and are thus key to
begin understanding the different results in poverty levels and trends as highlighted
above.
The common unit of measurement is the Purchasing Poverty Parities (PPP)
as established by the International Comparison Programme (ICP) of the World
Bank, a major statistical collaboration to collect prices of comparable goods in
countries around the world.12 In the most recent 2005 round of the ICP, price data
12 The reason market‐based exchange rates are not used to convert national currencies into a common unit of measurement is that these reflect demand and supply of currencies used in international transactions, and do not necessarily reflect differences in price levels especially for goods and services that are not subject to international trade. In labour surplus economies found in most developing countries such ‘non‐tradables’ are typically relatively cheaper than in more developed countries, and thus purchasing power for developing nations will tend to be understated using market exchange rates.
- 21 -
was collected in 146 participating countries, including 48 in Africa (World Bank,
2008).
Uganda participated for the first time in 2005 with UBOS collecting price data
for 683 items, amounting to 11,000 quotes per month, over a 12 month period
(personal communication with UBOS). Price points were in both urban and rural
areas but senior statisticians at UBOS confirm the “urban bias” also reported by
World Bank (2008) and Chen and Ravallion (2008) as some prices, especially for
non‐food items, are generally more readily available in urban areas. However, the
general perception is that the design and management, under the African
Development Bank for the Africa‐region, represents a significant improvement
compared to earlier rounds of the ICP and that the PPP values in Uganda are now
more reliable as they are based on “actuals” rather than a regression estimate.13
According to the documentation of the ICP, the 2005 round was “the most extensive
and thorough effort ever to measure PPPs across economies” (World Bank, 2008:9).
The second major step in establishing the IPL is to estimate one common
threshold that can be applied to country level PPPs. For this purpose Ravallion et al.
(2008) regresses per capita expenditure on the PPP value of poverty lines for 75
countries with a lower bound determined by the conditional mean value of the
poverty lines for the 15 poorest countries, of which Uganda is one.14 The
international poverty line is thus estimated at US$ 37.983 or the equivalent of US$
1.25/day. This figure is considerably higher than the 1993 estimate of US$1.08/day,
which adjusted for US inflation is equal to US$ 1.45/day in 2005. However, the
lowering of the international poverty line is not surprising given the general
downward revision of the PPPs for the poorest countries (including Uganda) that
was the outcome of the latest ICP round. The overall effect of the downward
13 PPPs for non‐benchmark countries (65 in 2005) are imputed based on an estimate of the PPP GDP/cap values of the benchmark countries regressed on GNI/cap and the secondary gross enrolment rate, both from the World Development Indicators (see World Bank 2008:164 for details). 14 The others are: Malawi, Mali, Ethiopia, Sierra Leone, Niger, Gambia, Rwanda, Guinea‐Bissau, Tanzania, Tajikistan, Mozambique, Chad, Ghana and Nepal. As only two countries are included from outside sub‐Saharan Africa, the global standard for absolute poverty is very much determined by this region.
- 22 -
revision in the poverty line has been a significant upward revision in the global
estimate of poverty from less than one billion in 2004 to 1.4 billion people in 2005.
The “dollar a day” poverty line continues to serve as one of the most
referenced development indicators and among its virtues is its simplicity,
applicability across countries and its anchoring in actual poverty lines of the poorest
countries (Deaton 2001). However, the international poverty line also continues to
be the subject of debate for its sensitivity to the revision of base year and the use of
PPP conversion factors that capture the price level of a general basket of goods and
not necessarily one consumed by the poor (Reddy and Pogge 2005; Deaton 2001).
Comparing poverty lines
Given these inherent differences in the two sets of poverty lines it is useful to
compare the implicit value of the two poverty lines for the two base years used in
Uganda. In order to achieve comparison there is a need to adjust for price changes
and the construction of the poverty lines (items 2 and 3 in Figure 4).
The procedure is to first convert the international poverty line of
US$1.25/day per capita estimated at the ICP base year ( $, ) into local
currency using the PPP exchange rate ( $, ) for that year as provided
by ICP (World Bank 2008). Then the local currency units (Ush) that are equivalent
to the IPL at each survey year, t, can be obtained by deflating using the consumer
price index (π), or:
, $, $, π
π
Table 5 contains the computations of this for the years 2005, 1997 and 1993.
It is possible to compare directly the level of the World Bank poverty line with the
national poverty lines for 1993 and 1997, the two years for which the Ugandan
poverty has been computed. The international poverty line in US$ PPP is converted
into Ugandan Shilling and then the national poverty line in local currency is
converted into US$ PPP using the same approach. It is clear that the Ugandan
poverty line equivalent is slightly higher but quite close to 1.25 in USD/PPP; Ush
- 23 -
21,091 compared to the 21,135 national poverty line. However, it is also clear that
using that when using 1993 as a baseline the national poverty line translates into a
higher poverty line in PPP values. The sensitivity to the choice of base year has been
discussed extensively in the context of the PPPs and the international poverty line,
and is a particular source of complication when comparing across countries or when
using different PPP estimates (Reddy and Pogge 2005; Deaton 2001). The
divergence in poverty lines in Uganda between 1997 and 1993 when converted
using the 2005 value is an indication that the World Bank and UBOS are using
different CPI estimates for the adjustments at least prior to 1997. However, with
1997 as the base‐year, differences are minor.
Table 5: Comparing the real values of the poverty lines
$, $, ,
Conversion from US$ PPP to local currency
Day
Month
2005 1.25 745 1.000 931 28,310
1997 1.25 745 0.745 694 21,091
1993 1.25 745 0.545 508 15,429
Conversion from local currency to US$ PPP
1997 1.25 745 0.745 695 21,135
1993 1.33 745 0.545 541 16,443
Sources: Authors calculations based on data from PovcalNet, Appleton (1999) and EPRC.
4.2 Adjusting for household composition
As noted above the common unit of analysis underlying both sets of poverty
estimates is household consumption expenditure. However, the two approaches
differ when it comes to adjusting for the composition of the household. In PovcalNet
household expenditure is divided by the number of household members so that
expenditure is expressed in per capita terms. UBOS uses a measure of adult
equivalents instead that weighs differently household members according to their
age. Full details of these weights are given in Appleton (1999) but the main feature
is that all persons under 18 are weighted less than one, (e.g. those less than 1 years
- 24 -
old are weighted 0.273 and those aged 16‐18 are weighted 0.95) and everyone
above 18 are weighted 1. The basic reasoning is that the consumption threshold of a
child is lower than of an adult e.g. as minimum required nutritional intake is lower.
This way among households of equal size and per capita expenditure those
households with the most and the youngest children will have a higher adult
equivalent income.
The impact of adjusting for adult equivalents is important for the poverty
estimates as is illustrated on Figure 5. The figure displays FGT‐curves for α = 0, or
cumulative density functions at different values of the poverty line, for the four
surveys, and for two measures of welfare; per capita consumption expenditure
(dashed curve) and adult equivalent consumption expenditure (solid curve) all in
1997 prices.
Figure 5: FGT curves (1997 prices)
Source: Author’s computations based on UBOS data.
Notes: Dashed curve represents household consumption expenditure per capita. Solid curve represents household expenditure adjusted for adult equivalents. A is the poverty line of USh 21135, B is the level of poverty when using adult equivalents, C is the level of poverty when using the per capita welfare measure and D is the level of poverty from PovcalNet.
A
B
C
D
0.2
.4.6
.81
FG
T(0
)
0 20000 40000 60000 80000 100000Poverty line (z)
1992/1993
A
B
C
D
0.2
.4.6
.81
FG
T(0
)
0 20000 40000 60000 80000 100000Poverty line (z)
1999/2000
A
B
C
D
0.2
.4.6
.81
FG
T(0
)
0 20000 40000 60000 80000 100000Poverty line (z)
2002/2003
AB
D
C
0.2
.4.6
.81
FG
T(0
)
0 20000 40000 60000 80000 100000Poverty line (z)
2005/2006
- 25 -
Results from the four surveys are presented in panels where line A marks the
official poverty line of 21,135 Ush/month in 1997 prices. In 2005/2006 this point
represents the 31.1 percent poverty estimate indicated by line B. Line C indicated
the level of poverty that is generated by using the welfare measure adjusted to
reflect per capita expenditure. Again for 2005/2006 the estimate is 49.9 percent,
which is very close to the PovcalNet estimate of 51.5 percent as indicated by line D.
For all years the Line C is above Line B, indicating that for a fixed poverty line the
household expenditure per capita will generate a higher poverty estimate than
expenditure per adult equivalent (in technical terms there is first order stochastic
dominance). The proximity of Line C to Line suggests that a key explanation for the
difference between the national and international poverty estimates lies in the
different ways that household composition is accounted for.
4.3 Data sources and sampling
While UBOS has sought to maintain comparability between the surveys over time,
invariably changes do occur and such changes need to be separated from actual
changes in economic and social welfare. These types of non‐sampling errors (item 8
in Figure 4) include: differences in the timing and duration of the surveys, the
coverage of the sampling frame, design of the survey instrument, geographical
coverage and panel elements (UBOS 2006).15 One specific challenge relate to the
representativeness of the household surveys. The sample design has broadly
remained the same since 1992/93 using a two‐stage stratification design. However,
the sampling frame was changed after the 2002 Census. With given rapid population
growth in Uganda some concern has been raised over the comparability of the
sampling frames. Also, implementation of the 1999/2000 survey suffered from
funding problems, which forced the sample size to be cut in the middle of the
15 UBOS uses the “recall method” to capture expenditure of households as compared to the diary method used in some countries. In the IHS of 1992/1993 the consumption module was open‐ended unlike the other surveys where potential items are listed and read by the enumerators, which is also the reason why the MS‐1 survey was used instead as a basis for the food basket used in the poverty line. The recall period varies according to the type of good with some changes between surveys. The consumption sub‐module has also been modified over time to include “new” areas of consumption such as mobile phones and to capture the same item in its different forms.
- 26 -
survey. To compensate for the change UBOS adjusted the sample weights and is
confident that the weighted 1999/2000 survey adequately represents the
population (World Bank 2006).
Moreover, with a high degree of insecurity in some parts of the country, not
all surveys have been able to cover the entire country. This remains a challenge for
understanding changes in poverty trends and typically poverty measures over time
will seek comparability by exclude the results from a number of districts (Gupta
2004; UBOS 2006). While UNHS‐1 had excluded the districts of Bundibugyo, Gulu,
Kasese and Kitgum, UNHS‐2 excluded certain parts of Gulu and Kitgum along with
Pader, a new district previously part of Kitgum. The estimates in Table 6 labelled
“Fixed sample” are run only using districts that were included in all the surveys.
When the 2002/2003 UNHS‐2 estimates are revised to exclude these districts from
the sample the incidence of poverty is revised down slightly from 38.8 to 37.7
percent while the revision for 2005/2006 UNHS‐3 implies a lowering from 31.1
percent to 28.9 percent. The official MDG Report seems to be inconsistent in its
reporting of poverty incidence data using “Fixed sample” estimates for the
1992/1993 IHS, UNHS‐1 and UNHS‐2 but not for the UNHS‐3 as reflected on Figure
1.
For purposes of future MDG Reporting it would be advisable to distinguish
between estimates according to which sample they are based on. When assessing
long‐term progress towards MDG attainment the relevant figures are from the “Full
sample” 56.4 percent in 1992/1993 and 31.1 percent in 2005/2006 or a reduction
of 45 percent. In other words the omission of the poorest districts in the analysis
increases mean consumption per adult equivalent and lowers the proportion of the
poor. However, it is not clear from the meta‐data of PovcalNet whether similar
adjustments have been made and thus whether comparability between survey data
from the different years has been established (item 6 in Figure 4). Nor is it clear the
extent to which the sample structure is taken into account to weight the data.
- 27 -
In Table 6 findings are also presented for poverty estimates using just one
national poverty line instead of the 8 regional ones.16 There is very little difference
in the estimates of poverty, maximum 1 percentage point, and none of these
differences are statistically significant. This particular aspect of the construct of the
poverty line (item 2 in Figure 4) is thus of limited impact on the diverging poverty
estimates. These findings are robust to the survey year selected, the choice of
sample and irrespective of whether or not adjustments are household composition.
This is in line with findings by Appleton (2003) for the earlier years.
Table 6: Regional and national poverty lines
Whole sample 1992/93 1999/00 2002/03 2005/06
Regional poverty lines
56.4% 33.8% 38.8% 31.1% (54.2%‐58.6%) (31.6%‐36.0%) (36.8%‐40.8%) (29.2%‐33.0%)
National poverty Line
57.4% 34.8% 39.6% 31.5% (55.2%‐59.6%) (32.6%‐37.0%) (37.5%‐41.6%) (29.5%‐33.5%)
Fixed sample 1992/93 1999/00 2002/03 2005/06
Regional poverty lines
55.7% 33.8% 37.7% 28.9% (53.4%‐58.0%) (31.6%‐36.0%) (35.6%‐39.7%) (26.9%‐30.8%)
National poverty line
56.6% 34.8% 38.5% 29.3% (54.3%‐58.9%) (32.6%‐37.0%) (36.4%‐40.6%) (27.2%‐31.3%)
Notes: “Whole sample” includes all districts covered in the survey that year whereas ”Fixed sample” only includes districts that were covered in all the surveys, i.e. excluding Bundibugyo, Gulu, Kasese, Kitgum and Pader.
Source: Author’s computations based on UBOS data.
4.4 The role of price adjustments The final set of issues that will be explored in this paper relates to the role of prices
in estimating poverty incidence (item 11 in Figure 4). As explained above the
measure used by World Bank and PovcalNet adjusts the international poverty line
by the PPP conversion factor, which is deflated using consumer price indices. The
measure of household consumption is converted using the real PPP values. In
16 For purpose of calculations in this paper the difference is that under the regional poverty line estimates the household consumption expenditure is normalised using the ratio of the regional poverty line to the national poverty line. For the national poverty line estimates, the welfare variable was not normalised. The same national poverty line of Ush 21135 was used in both cases. The values of the regional poverty lines are listed in Table 4.
- 28 -
Uganda the adjustment process is a bit more elaborate. Two sets of adjustments are
made. The first is infrequent and, as mentioned earlier, is to rebase the poverty line
using the national CPI to deflate the food and non‐food components separately.
The second set of adjustments takes place after every new survey. Three
conversions are made to express the consumption aggregate in constant prices:
1) Revaluation of home food consumption into market prices using the
median unit values from the household consumption data. This is done separately
for the four administrative regions sub‐stratified into rural and urban, making 8
sub‐regions.
2) Spatial price variations, which are found to be significant both within and
across regions, are accounted for by using price ratios according to the same 8 sub‐
regions to the national weights of the CPI.
3) Inter‐temporal price variations are accounted for by adjusting the
consumption aggregate for inflation using the average all‐item CPI for the survey
period but without taking into account the particular month a household was
interviewed.17
In the application of 3), World Bank/PovcalNet and UBOS share the approach
although it is a problematic one that neglects possible differences in the changes in
prices according to regions and zones. As shown by Appleton and Ssewanyana
(2003) these differences can be significant and have a distinct pattern of faster
increases in prices in urban areas. In this situation using CPI to update the poverty
line and the consumption aggregate will underestimate the purchasing power of the
rural areas, where most of the poor live. In turn, this will tend to exaggerate poverty
levels. This is well‐acknowledged but UBOS chooses to use the CPI adjustments
because of its more regular production rather than use less‐frequent survey based
information. For PovcalNet/World Bank it is also a matter of convenience and
standardisation given data availability. Since UBOS and PovcalNet use a similar
17 Consumption expenditure is deflated differently according to the recall period: 7‐day and 30‐day recall items are deflated using the average of CPI in the current and previous months for all the survey period and taking the average; whereas 365‐day recall items are deflated using the average CPI for the survey period and the equivalent of survey period before the survey (Ssewanyana and Muwonge, 2004).
- 29 -
approach in accounting for inter‐temporal price changes it is less likely that this is a
source of diverging poverty estimates. Obviously this is dependent on the use of
comparable CPI numbers, which has not been possible to verify for either of the two
sources, although the discussion above suggested that this might be a problem at
before 1997.
Figure 6: Adjustments to household consumption expenditure
Sources: UBOS (2003, 2006); Appleton (1999).
The two first adjustments are also potential sources of divergence in as much
as one or both are not taken into account in the PovcalNet data. Both types of
adjustments tend to have a positive impact on household expenditure in rural areas.
The combined upward adjustment for rural areas is around 10 percent where it is
typically negative for urban households as depicted on Figure 6. The effects of
leaving out these adjustments would be to lower the absolute and relative level of
rural household incomes and raise the overall level of poverty. However, given the
lack of detailed information information on the consumption aggregate it is not
possible at this stage to determine with certainty whether differences in the price
adjustments can further explain the unexplained elements of the diverging poverty
- 30 -
estimates. However, given the mismatch in values between the per capita means of
household consumption reported in PovcalNet and those reported by UBOS these
are likely factors.
4.5 Summary of effects of underlying causes
In Table 7 all the likely underlying causes of divergence in the poverty estimates are
summarised. The starting point in (line a) are the estimates from the MDG Report
followed by the changes in that estimate that is attributable to each of the
adjustments discussed above (lines b‐f). Finally, all these adjustments are combined
to produce poverty new estimates with (line g) and without the sampling weights
(line h). The net result from this ‘peeling off’ from the national poverty estimates a
series of standard and highly relevant adjustments is quite dramatic and makes the
new estimates move quite close to those of PovcalNet. From the outset the national
estimates were anywhere from 56 percent to 80 percent of the international ones.
However, the adjustments made bring the two sets of estimates into a 5‐6 percent
difference. It is particularly interesting to note that between line g and h, i.e. through
the un‐weighting of the estimates, the increase in poverty between 1999/2000 and
2002/2003 disappears, just as it does in the estimates from PovcalNet.
Table 7: Difference in poverty estimates
1992/1993 1999/2000 2002/2003 2005/2006
a. MDGR/UBOS estimates 0.557 0.338 0.388 0.311
b. Full sample 0.564 0.338 0.388 0.311
c. National poverty line 0.574 0.348 0.396 0.315
d. Lower value of poverty line 0.563 0.337 0.386 0.309
e. Per capita 0.739 0.547 0.589 0.499
f. Price adjustments 0.658 0.405 0.424 0.352
g. All adjustments (b-f) 0.801 0.607 0.620 0.541
h. All adjustments + un-weighted 0.741 0.578 0.548 0.539
i. PovcalNet 0.700 0.605 0.574 0.515
j. Difference before adjustments (a/i) 80% 56% 68% 60%
k. Difference after adjustments (h/i) 106% 95% 96% 105%
Source: Author’s computations based on UBOS data.
- 31 -
4.5 Metadata: issues in access and quality
A final word before concluding needs to be said about the meta‐data for the two
sources since limitations, especially in the data sources in PovcalNet, has
complicated this exercise and made the investigation into the underlying causes
quite speculative. On the side of UBOS the meta‐data is quite strong, and this is
probably facilitated by the implementation of the Uganda National Data Archive
(UNDA), which provides a catalog of surveys, their meta‐data and the data itself on a
web‐based platform. A total of 10 Censuses and surveys are currently covered
including the four UNHS used in this paper. Information on the computation of the
poverty line, and the different adjustments made to the welfare variable, is partially
available in research papers and reports from UBOS and EPRC but could in the
future be added to the UNDA or another depository in UBOS to increase
transparency.
Figure 7: Access to survey information in PovcalNet
Source: PovcalNet online version accessed June 2010..
The meta‐data from PovcalNet is less complete. According to PovcalNet the
household survey data comes from “World Bank country teams”, and the meta‐data
section on MDG1/1 from the UN Statistics Division reads: “Only nationally
representative surveys that are of good quality, contain sufficient information to
produce a comprehensive consumption or income aggregate, and allow for the
- 32 -
construction of a correctly weighted distribution of per capita consumption or
income are used.”18 Under meta‐data section the issue of missing values is referred
to only in relation to yearly observations of the poverty indicator, which “is
calculated only in years and for countries for which suitable survey data are
available.”
It is unclear which criteria the World Bank and the UN Statistics Division has
established to determine whether a survey is “of good quality” and why it is the
“World Bank country team” and not the National Statistics Office directly that
supplies the data (item 10 in Figure 4). In the information that accompanies country
computations in PovcalNet19 UBOS is listed under “Survey Conductor” and “Data
Access”. In this context it is a bit puzzling that the World Bank has decided to
include the HBS from 1989 in PovcalNet, and have thus found it to be of “good
quality”, when this survey is no longer used by UBOS and researchers for its lack of
comparability with other surveys.20 The meta‐data section of PovcalNet holds
additional pieces of information that creates uncertainty about what is included in
the data base. For instance the sampling size of MS‐2 and MS‐3 is stated as 5,000
even as UBOS/EPRC report sampling sizes of 4,925 and 5,515 respectively
(Ssewanyana and Muwonge, 2004). PovcalNet also refers to a National Integrated
Household Survey of 1996 with a sample of 29,745 but according to UBOS the only
household survey conducted in 1996 was MS‐3 with a sample size of 6,654. As
depicted on Figure 7, users of PovcalNet are not authorised to access the specific
information regarding the surveys, which is problematic in terms of transparency.
Moreover: the section in the meta‐data file on the CPI used for adjustment is
missing; information about the sample design is limited and not up to date; the
section 6 on “Documentation” is blank, and; under the subsection on “Details of
consumption and/ income aggregates” the information is noted as “N.A.” As
18 http://mdgs.un.org/unsd/mdg/Metadata.aspx?IndicatorId=1 (accessed June 2010). 19 http://iresearch.worldbank.org/PovcalNet/doc/UGA.htm#3 (accessed June 2010). 20 Appleton (1999) points to overestimation of consumption in the HBS and attributes this to sampling issues as mean household size was a full one person higher in the HBS than in the census in 1991 and in the subsequent household surveys.
- 33 -
mentioned above it has also not been possible to find information regarding the
application of probability weights to the survey sample data.
5. Conclusions
This paper has explored differences in the poverty estimates from Ugandan
authorities and the World Bank. These differences have been linked to proximate
causes in the specification mean level of welfare and the distribution of that welfare.
The underlying causes for the divergence appear to be attributable to differences in
adjustments of price variability between urban and rural areas, adjustments to
household composition and the probability weights associated with the sampling
structure. But issues related to transparency of the meta‐data especially when it
comes to the PovcalNet have also been raised.
Since the national poverty estimates are constructed to reflect local
circumstances such as spatial price differences and the consumption patterns of the
poor, as well as the country and time‐specific composition of households, the
measure used by UBOS and used for national MDG reporting measure is clearly
favoured over the “dollar a day” measure for measuring national level poverty and
for tracking progress towards national and international development goals. The
adjustments made by national authorities to the Uganda data and the way that the
poverty line is derived follow standard practice in poverty measurement. When
assessing the overall question as to whether Uganda is on track to meet MDG1/1 the
answer is thus clearly in the affirmative given the rapid downward trend in poverty
trends as expressed by the national poverty line. This is also corroborated by other
types of data that reflect other dimensions of welfare and methodological
approaches such as those related to household assets and physical features of the
households (UBOS 2006; Younger 2003).
Nevertheless, as this paper has pointed to there are a series of opportunities
for strengthening the way national and international agencies measure and report
poverty estimates. The key recommendations of this study are:
- 34 -
1. The World Bank’s PovcalNet is a great innovation and a tremendous
source of information on poverty in the developing world. However, the
preceeding analysis has highlighted the potential for strengthening the
meta‐data section of PovcalNet in particular so that it is clear which
surveys have been included and why, what if any adjustments have been
made and some notes on comparability e.g. as sampling changes over
time. Better sourcing of information on the applied PPP conversion
factors would also be useful. This could be achieved by completing the
missing information in the already existing meta‐data features of
PovcalNet, granting user access to the currently restricted areas that
holds information about each of the survey and/or adding new
information. Insufficient documentation in secondary databases of the
World Bank is not a new issue (Pyatt 2003; Székeley & Hilgert 1999) and
this paper finds applies to PovcalNet as well.
2. The commitment by the World Bank to publish national and international
poverty estimates side by side in its flagship publications is noteworthy
but follow‐through needs to be consistent as the example from Uganda
clearly shows. Moreover, a real question is whether it is sufficient to just
publish the numbers together or whether efforts should be made to
include a clear statement of guidance for how to interpret and compare
the numbers. Country‐level debates can be frustrated (as has been the
case in Uganda) when there is ambiguity in the numbers, and official data
can lose credibility when viewed next to very different numbers from
reputable institutions such as the World Bank and the United Nations.
3. In its current form PovcalNet allows users to change only the level of
poverty line. Future versions might add more options for users, to test for
sensitivity of results including for differences in household composition.
As the availability of unit data expands this should become increasingly
feasible. As this study has documented, whether or not probability
weights are applied to the sample survey data influences the estimates
quite profoundly and this issue needs to be addressed in PovcalNet.
- 35 -
Obviously none of these suggestions for improvements will address the
fundamental criticisms with regards to the setting of the international
poverty line, its sensitivity to choice of base year and the use of PPP
conversions, which have led some to suggest that the dollar‐a‐day
poverty measure be abandoned altogether and alternate approaches for
measuring global poverty be established (Pogge and Reddy 2005; Deaton
2001). Although some divergence is unavoidable proposals for
establishing new global measures of poverty will be particularly helpful if
they generate estimates that do not contradict national level measures.
4. At the national level, this study has shown MDG progress reporting in
Uganda should be mindful that only comparable poverty estimates are
used, notably considering the differences in sampling in the 1999/2000
survey. The practice in the 2007 MDG report of inconsistent reporting
should be avoided. One way would be to present the two sets of estimates
for all districts (full sample) and those that are common in all the surveys
(fixed sample) side by side. With the acceleration of economic and social
development in northern Uganda in recent years including these districts
in poverty trends and analysis is essential.
5. UBOS is considering a revision of its poverty line due to changes in
consumption patterns. In the likely event that such a revision takes place
it would be advisable to continue computing poverty estimates based on
the original set of expenditure weights. This would facilitate
comparability and given the baseline proximity to 1990 this would also
facilitate MDG reporting. However, national authorities need to be
mindful of issues related to welfare consistency in comparison of poverty
over time should the poverty line be changed to take into account
changing patterns of consumption. Two additional issues should also be
considered. Firstly, the food poverty line could be allowed to vary across
regions to reflect the large regional variations in diet, the cost
effectiveness of different stables in providing calories, and regional
differences in income levels. Such an exercise would be particularly
- 36 -
interesting given the insignificant differences in poverty estimates that
was found in this paper between the application of the current 8 regional
poverty lines (that reflect only variation in the non‐food component of
the poverty line) or just one national line. Secondly, an upper bound
poverty line could be included whereby non‐food expenditure are
estimated at the level of total food expenditure of households around the
food poverty line, in comparison with current more austere definition of
using total expenditure. Sometimes countries report both types of poverty
lines side by side as to represent different severities of poverty. Finally,
Uganda is in the process of implementing an overdue collection of
consumer price data in rural areas, which will be essential to better
capture the heterogeneity of changes in price levels.
6. Greater transparency is needed in the procedure for computing national
level poverty measures, rebasing the poverty line and adjusting
consumption expenditure to fixed‐price units. While much of this
information is available in a series of excellent reports and papers, and
through support from officials and researchers working on the household
data, there is a need to systematically collect, file and make available this
information.
7. For purposes of facilitating comparability of poverty estimates in the
longer term, greater efforts should be made in the African region and sub‐
regions for coordinating the compilation of poverty statistics and for
harmonising methodologies and survey cycles. This would facilitate
future cross‐country comparability in an area where little exists today.
National statistics offices and international development partners such as
the UN and the World Bank should play key roles in coordinating such an
initiative.
Ultimately, even if the national and international poverty estimates serve
different purposes, it should be expected that they broadly reflect similar levels and
changes in welfare. As this paper has shown, this is clearly not the case in Uganda.
- 37 -
That should not take away from the great achievements the country has made nor
the challenges it has faced along the way and the challenges that remain.
6. References
Araar, A. & J‐Y. Duclos (2007) “DASP: Distributive Analysis Stata Package.” Quebec: Universite of Laval.
Appleton, S. (2003). “Regional or National Poverty Lines? The Case of Uganda in the 1990’s.” Journal of African Economies, 12(4):598‐624.
Appleton, S. (1999). “Changes in Poverty in Uganda 1992‐1997.” Working Paper Series/99‐22 Centre for the Study of African Economies, University of Oxford.
Appleton, S. (1998). “Changes in Poverty in Uganda 1992‐1996.” Working Paper Series/98‐15. Centre for the Study of African Economies, University of Oxford.
Appleton, S. and Ssewanyana, S. (2003). “Poverty estimates from the Uganda National Household Survey II, 2002/2003.” Unpublished paper, University of Nottingham and Economic Policy Research Center.
Chen, S. and Ravallion, M. (2008). “The Developing World is Poorer than We Thought, But No Less Successful in the Fight Against Poverty.” Policy Research Working Paper No. 4703. Washington DC: World Bank.
Deaton, A. (1998). “The Analysis of Household Surveys: A Micro‐Econometric Approach to Development Policy.” Baltimore and London: World Bank and Johns Hopkins Press.
Deaton, A. (2001). “Counting the World’s Poor: Problems and Possible Solutions.” World Bank Research Observer, 16(2): 125‐127.
Datt, G. (1998) “Computational Tools for Poverty Measurement and Analysis”, IFPRI Food Consumption and Nutrition Division Discussion Paper No. 50. Washington: International Food Policy Research Institute.
Dijkstra, A. G. and van Donge, J. K. (2001). “What does the ‘‘Show Case’’ Show? Evidence of and Lessons from Adjustment in Uganda.” World Development 29(5): 841–63.
Foster, J., Greer, J. and Thorbecke, E. (1984). “A Class of Decomposable Poverty Measures.” Econometrica, 52: 761‐6.
Gupta, S. K. (2004). “Uganda National Household Survey Programme. Interim Report on Comparison of UNHS 1999/00 and 2002/03 Results.” Unpublished paper, July 2004.
Kappel, R., Lay, J. and Steiner, S. (2005). “Uganda: No More Pro‐poor Growth?” Development Policy Review, 23(1):27‐53.
Lanjouw, P. and M. Ravallion (1995). “Poverty and Household Size.” Economic Journal, 105(433): 1415‐34.
- 38 -
Luoto, J. (2007). “What Happened in the UNHS2?: Re‐estimating Uganda’s Poverty Numbers.” Unpublished paper, University of California, Berkely.
Lustig, Nora (2007) “Data Availability and Consistency for MDG Monitoring and Reporting: an Assessment.” Report to the UNDP Poverty Group, July 23rd 2007 revision, mimeographed.
Mallaby, S. (2004). “The World’s Banker: A Story of Failed States, Financial Crises and the Wealth and Poverty of Nations.” Penguin Press, New York.
Mackinnon, J. and Reinikka, R. (2002). “How Research can Assist Policy: The Case of Ugandan Economic Reforms.” World Bank Research Observer, 17:267‐92.
McGee, R. (2004). “Constructing Poverty Trends in Uganda: A Multi‐Disciplinary Perspective.” Development and Change 35(3):499‐523.
Ministry of Finance, Planning and Economic Development (2004). “Poverty Eradication Action Plan 2004/05–2007/08.” Kampala: Government of Uganda.
Okidi, J., Ssewanyana, S., Bategeka, L., and F. Muhumuza (2004) “Operationalising Pro‐Poor Growth. A Country Case Study on Uganda.” A joint initiative of AFD, BMZ (GTZ, KfW Development Bank), DFID, and the World Bank.
Pyatt, G. (2003). Development and the Distribution of Living Standards: A Critique of the Evolving Data‐base.
Ravallion, M. (2002). “How Not to Count the Poor: A Reply to Reddy and Pogge.” Mimeographed, Development Research Group, World Bank.
Ravallion, M. (1998). “Poverty Lines in Theory and Practice.” LSMS Working Paper No 133. Washington DC: World Bank.
Ravallion, M. and Bidani, B. (1994). “How Robust is a Poverty Line?” World Bank Economic Review 8(1): 75‐102.
Ravallion, M., Chen, S. and Sangruala, P. (2008). “Dollar a Day Revisited.” Policy Research Working Paper 4620. Washington DC: World Bank.
Reddy, S. and Pogge, T. (2005). “How Not to Count the Poor.” Mimeographed, Columbia University.
Szekely, M., Lustig, N., Cumpa, M. and Mejia, J. A. (2004). “Do We Know How Much Poverty There Is?” Oxford Development Studies, 32(4): 33‐56.
Szekely, M. & M. Hilger. (1999). “What’s Behind the Inequality we Measure: An Investigation Using Latin‐American Data.” Inter‐American Development Bank , 32(4): 33‐56.
Ssewanyana, S. and Muwonge, J. “Measuring and Monitoring Poverty: The Uganda Experience.” Paper prepared for the Poverty Analysis and Data Initiative (PADI) meeting, Mombasa, Kenya, May 6‐8, 2004.
Ssewanyana, S. and Muwonge, J. “Prospects and Challenges of Measuring Poverty in Uganda.” Paper presented at the Scientific Statistics Conference, Kampala, Uganda, 11‐13 June, 2007.
- 39 -
Ssewanyana, S. and Okidi, J. (2007). “Poverty Estimates from the Uganda National Household Survey III, 2005/2006.” Occasional paper No 34. Kampala: Economic Policy Research Center.
Uganda Bureau of Statistics, “Uganda National Household Survey 2005/2006: Report on the socio‐economic module.” December 2006. Kampala: Uganda Bureau of Statistics.
Uganda Bureau of Statistics, “Uganda National Household Survey 2002/2003: Report on the socio‐economic module.” November 2003. Kampala: Uganda Bureau of Statistics.
Uganda Bureau of Statistics, “Uganda National Household Survey 1999/2000: Report on the socio‐economic module.” Kampala: Uganda Bureau of Statistics
UN (2008). “The Millennium Development Goals Report 2008.” New York: United Nations.
UN (2007). “Millennium Development Goals: Uganda’s Progress Report 2007.” Kampala.
UN (2003). “Indicators for Monitoring Millennium Development Goals.” New York.
UN (2001). “Road Map Towards the Implementation of the United Nations Millennium Declaration.” Fifty Sixth Session of the UN General Assembly. A/56/326, 6 September.
UN Development Group (2003). “Country Reporting on the Millennium Development Goals. Second Guidance Note.” Accessed Nov 2008 at http://www.undg.org/.
World Bank (2008a). “World Development Report 2008.” Washington DC: World Bank.
World Bank (2008b). “Global Purchasing Power Parities and Real Expenditures: 2005 International Comparison Program.” Washington DC: World Bank.
World Bank (2006). “Uganda Poverty and Vulnerability Assessment.” Washington DC: World Bank.
Younger, S. (2003). “Growth and Poverty Reduction in Uganda, 1992‐1999: A Multi‐Dimensional Analysis of Changes in Living Standards.” Cornell Food and Nutrition Policy Program Working Paper No 151.